Who runs this, and by what rules

Why trust a site run by one person

You should not have to. The site is built so its claims can be checked without trusting its author: every number traces to resolved on-chain events anyone can verify, the methodology is published in full with its limitations, the alert record grades itself in public with losses included, and the validation report was committed to publication before its results existed. The receipts ledger publishes its own arithmetic identity, so silent edits would be detectable by any reader.

Corrections

Errors are corrected the same way alerts are graded: in public. When a published number or graded row is found to be wrong, the correction is applied and noted on the affected page rather than silently swapped, and anything material enough to change a conclusion is called out where the conclusion was drawn. Grading itself is mechanical: alert rows are recorded at detection, frozen, and settled by each market’s recorded resolution, with no editing after the fact. Voided markets are excluded from the record and labeled as such. Every correction is logged below.

Correction log

Every published number, verdict, or claim found to be wrong, newest first, including the ones nobody outside would have noticed. A product that grades other people’s records in public should be able to show its own.

• 2026-08-05: Receipts ledger: re-alert snapshots were counted as separate rows. Detection re-alerts a position as the tracked wallet adds to it, and each snapshot carries the position's cumulative size. The ledger recorded every snapshot as its own row, so one growing position could appear several times: 1,430 rows over 644 distinct positions. Re-alerted positions won more often than average, so row counting flattered the record: 723 won to 379 lost by rows (66% of resolved), 344 won to 209 lost by positions (62%). The ledger now carries one row per position, keeping the first alert's date, price, and the market prices 1 and 24 hours after it, the largest size observed, and an n_alerts column in the public JSON ledger stating how many times the feed flagged the position. Individual grades were never wrong; every snapshot of a position grades identically. The same double counting inflated the daily email and Telegram digest's header totals, fixed the same day. Figures are from the 2026-08-05 verification run.
• 2026-07-26: Win-rate guide: skill-prevalence shares replaced after the hedging correction. The guide published the share of records statistically distinguishable from luck as 6%, 27%, and 45% by evidence depth. Two things were wrong with that. The test behind it counted hedged and traded-out events as forecasts, which inflates thin records, and the shares were computed as the ahead-of-prices tail alone while the wording did not say so. On the corrected test, which grades only positions held to resolution on one side, the same ahead-tail statistic reads 1.3%, 4.5%, and 44.5%, and the page now states both tails: 12.0%, 29.8%, and 22.4% of records are distinguishable and behind their prices. The graded population moved from 640,363 to 377,736 wallets. Our interim note predicted the figures would fall substantially; that held for the thin tiers, and the deepest tier barely moved. Every figure on the page comes from the 2026-07-26 verified run, and distinguishable shares are stated by direction everywhere they appear.
• 2026-07-25: Wallet checker: hedged positions were graded as forecasts. The checker's test treated each event as a single bet at the price paid. Where a wallet held both sides of an event, or entered and exited before it resolved, that assumption was wrong, and wrong in one direction: a pair bought at 40 cents a side profits whichever way the event lands, so arbitrage and profit-taking were scored as forecasting accuracy. 39% of all wallet-events carried that signature. The checker now grades only positions held to resolution on one side and says so on every result. The skill-prevalence shares on the win-rate guide came from the same test and are marked under revision until the next verified data run replaces them.
• 2026-07-25: Wallet checker: a record behind its own prices could read as clearing the bar. The significance test flags a record as distinguishable in either direction, better or worse than the prices it paid. Every surface rendered that as a positive result, and the checker attached a subscription prompt to it. Wording and next steps are now split by direction, and a record that resolved behind its prices never receives the positive framing.
• 2026-07-25: Wallet checker data was four days old behind a current timestamp. The table behind the checker was published to the live site but was never included in the nightly rebuild, so it carried figures from 21 July while the page reported a current as-of date. It is rebuilt nightly now, and an automated check fails the build if any published table is missing from the rebuild.
• 2026-07-25: Statistics page figures were ten days old. The graded-alert counts read 745 sent and 517 resolved while the live ledger showed 1,043 and 726. Each block on that page now carries the date of the data run that produced it, so one stale section cannot sit behind a page-level date.
• 2026-07-24: Benchmark: a claim about political markets was not supported by the data. The calibration benchmark said our politics figures showed the price compression reported in the academic literature. A category average cannot demonstrate that, and the underlying price-bin curve ran the other way at short horizons. The claim was replaced with what the data shows, and the supporting table was published alongside it.

Methodology changes

The scoring method is versioned and dated on the methodology page. When the method changes in a way that moves rankings, the change is recorded there with its date and reasoning; historical grades on the receipts ledger are never rescored retroactively.

Contact

Questions, corrections, or data requests: hello@overround.pro. Data journalists and researchers can quote the statistics page with attribution; the public datasets are licensed CC BY 4.0.